Efficient MRI Reconstruction with Reinforcement Learning for Automatic Acquisition Stopping

Magnetic resonance imaging (MRI) is accelerated through subsampling of the associated Fourier domain in current clinical practice. The decisions on subsampling strategies and acceleration factors are provided heuristically before the acquisition. In this paper, we propose a reinforcement learning strategy for automatically deciding a subsampling strategy and acceleration factor for cardiac image acquisition. We build an environment that has a set of actions, including which k-space line to select next and when to stop the acquisition. We propose to use a reward term that penalizes extra line acquisitions and favours improved image quality. Experiments on cardiac MRI with different weightings of the reward function have shown that our method can achieve better image quality results without increasing the acquisition time and can automatically stop the k-space sampling process.